The Kinetic Limit of Balanced Neural Networks

TL;DR

The paper rigorously derives the kinetic (large-$n$) limit of a high-dimensional, balanced neural network with excitatory and inhibitory populations, nonlinear intrinsic dynamics, multiplicative noise, and spatially structured connectivity. By projecting dynamics onto a balanced manifold and decomposing into mean and fluctuations, it obtains autonomous equations for the mean activity and a Fokker-Planck description for fluctuations; with a Gaussian neural-field assumption, the limit further reduces to a low-rank neural-field system describing spatially distributed activity and covariances. The results are validated numerically in mean-field and nonlinear scenarios, showing good agreement with stochastic simulations in inhibition-stabilized and nonlinear gain regimes, while highlighting limitations and oscillatory regimes under spatial connectivity. The work provides a rigorous framework for understanding how large balanced networks maintain stability and how spatial structure shapes population variability, enabling reduced models that preserve essential dynamics. Overall, the paper advances the theoretical understanding of kinetic limits in balanced networks and lays groundwork for exploring spatiotemporal patterns within a principled, low-dimensional description.

Abstract

The theory of Balanced Neural Networks is a very popular explanation for the high degree of variability and stochasticity in the brain's activity. Roughly speaking, it entails that typical neurons receive many excitatory and inhibitory inputs. The network-wide mean inputs cancel, and one is left with the stochastic fluctuations about the mean. In this paper we determine kinetic equations that describe the population density. The intrinsic dynamics is nonlinear, with multiplicative noise perturbing the state of each neuron. The equations have a spatial dimension, such that the strength-of-connection between neurons is a function of their spatial position. Our method of proof is to decompose the state variables into (i) the network-wide average activity, and (ii) fluctuations about this mean. In the limit, we determine two coupled limiting equations. The requirement that the system be balanced yields implicit equations for the evolution of the average activity. In the large n limit, the population density of the fluctuations evolves according to a Fokker-Planck equation. If one makes an additional assumption that the intrinsic dynamics is linear and the noise is not multiplicative, then one obtains a spatially-distributed neural field equation.

Figures (5)

• Figure 1: Time evolution of the excitatory (blue) and inhibitory (red) mean population activities. Solid lines show the deterministic predictions $v_e(t)$ and $v_i(t)$ from \ref{['eq:ODEvs41']}, while the dotted lines plot empirical averages from the stochastic simulation. The parameters are $n=40000$, $\tau_e = 1$, $\tau_i = 1$, $\sigma_e=1$, $\sigma_i=1$, $A_e=1$, $C_{ei} = 1$, $C_{ie} = 1$, $C_{ii} = 0.5$, $G_{ei}(z) = z$, $G_{ie}(z) = z$, $G_{ii}(z) = z/2$.
• Figure 2: Fluctuation variances of the excitatory ($K_e(t)$, blue) and inhibitory ($K_i(t)$, red) populations. Solid curves are the analytic solutions of \ref{['eq:ODEKs41']} started at steady‐state $\tau\,\sigma^2/2$, and dotted curves show the sample variances from the stochastic simulation. The parameters are the same as in Figure \ref{['fig:mean_activity']}.
• Figure 3: Dynamics of the excitatory (blue) and inhibitory (red) mean activities under the nonlinear gain $G_{\alpha\beta}(z)=C_{\alpha\beta}\tanh(z)$. Solid curves show the deterministic predictions $v_e(t)$ and $v_i(t)$ obtained by solving the four‐variable ODEs from \ref{['eq:ODEvs42']}, while dotted curves trace the empirical averages from the stochastic simulation ($n=10\,000$, $\tau_e=\tau_i=1$, $\sigma_e=\sigma_i=1$, $A_e=0.1$, $C_{ei}=C_{ie}=1$, $C_{ii}=0.5$).
• Figure 4: Evolution of the fluctuation variances $K_e(t)$ (blue) and $K_i(t)$ (red) under the same nonlinear‐gain setup. Solid lines are the analytic variance ODE solutions of \ref{['eq:ODEKs41']} with initial conditions $K_{e}(0)=1$, $K_{i}(0)=2$, and dotted lines are the sample variances from the stochastic simulation. The parameters are the same as Figure \ref{['fig:mean_activity42']}.
• Figure 5: Stochastic evolution of the average excitation and inhibition. On the left $n=100$, and on the right $n=500$. The other parameters are $\tau_e = \tau_i = 0.5$, $\sigma_e = \sigma_i = 0.5$, $c_{ee} = [0.5,2]$, $c_{ei} = [4,4]$, $c_{ie} = [1,2]$ and $c_{ii} = [1,2]$. There does not seem to exist a balanced state in these simulations. As $n\to\infty$, oscillations arise, and the frequency increases with $n$.

Theorems & Definitions (20)

• Remark 2.3
• Lemma 3.1
• Theorem 3.2
• Corollary 3.3
• Proof 1
• Lemma 3.4
• Proof 2
• Corollary 3.5
• Proof 3
• Lemma 6.1